Skip to content
Explanation

The neutral verification layer

Skills became packaged software, and a software supply chain needs three answers before anyone trusts it: who signed the package, what it is allowed to do, and whether it still works after the model changes. This page is why we think an open body should give all three.

Agent skills became packaged software: versioned, scripted, installed from third parties, running inside your agent's privileged context. A software supply chain needs three answers before anyone trusts it — who signs the package, what it is allowed to do, and whether it still works after the model changes. Today, no neutral, open body answers any of the three.

Centralised players answer some of it behind their platform. Model vendors answer it only for their own harnesses. Nobody answers the question that actually hurts teams: "a model update shipped last Tuesday — do my 40 skills still pass?"

SkillProof is the fixed version of the agent-skill registry idea. We do not host skills, do not build new signing infrastructure, and do not sell policy enforcement. We host a tiny, signed, content-addressed ledger of attestations: for any skill at any content hash, what capabilities it uses and whether its eval pass-rate survived the last model bump. Any registry, any harness, any CI can consume it.

We are not npm. We are the thing npm never had and had to retrofit — the neutral verification layer.

Neutrality is the only durable position. Registries can display our attestations; scanners can consume our manifests; harnesses can gate on our diffs. A thing everyone can use and no one has to pay for spreads faster than a thing that requires a purchase order. If the big players absorb the standard, that is how open infrastructure wins — even when the company doesn't.

Core is MIT, forever. The ledger is public, forever. Everything here is falsifiable — including the decision to build it.

Where this stands today ​

This page describes the position, not the release. The commands exist, the schemas exist, and the ledger they are meant to feed does not; signing and the distributed eval harness are still ahead. The roadmap separates what is built from what is argued for.

Proof-first skill management for agent-first workflows